Chap e 27
A Sel -Adap i e ML Pipeline
o Sus ainable Manu ac u ing
Ramon Angos o A igues , And ea Fe nández Ma ínez ,
And ea G ego es Co o , and Jona han Josue To ez He e a
Abs ac The Eu opean indus y win g een and digi al ansi ion is c i ical o os e
sus ainable manu ac u ing p ac ices ha minimize and p e en he en i onmen al
impac and esou ce exhaus ion, while ensu ing Eu opean global compe i i eness. AI
can play a pi o al ole in his ansi ioning helping o op imize esou ces, enhancing
e iciency, and educing was e. None heless, AI adop ion in he indus y is s ill limi ed
due o he skills gap ha hinde s he managemen o hese ad anced sys ems. This
pape p oposes a sel -adap i e pipeline o ML wi h he inco po a ion o a new
O e i ing Index (OI) o sel -pa ame e uning emphasizing o e i ing p e en ion.
The pipeline inco po a es sel -pa ame e explo a ion capabili ies exploi ing su o-
ga e models o imp o e compu a ional e iciency. The p oposed OI o ML unc ion
is e alua ed in a well-known eg ession p oblem, p io o i s eal e alua ion in an
aluminum ecycling applica ion o suppo ope a o s selec ing he bes combina ion
o sc aps. Resul s indica e ha con igu a ions wi h lowe OIs demons a e supe io
gene aliza ion and obus ness, wi h he su oga e model e ec i ely iden i ying and
e ining high OI con igu a ions. The p oposed me hodology p o ides a baseline o
u u e de elopmen and in eg a ion o sel -adap i e and sel -imp o ing ML solu ions
in he indus y.
Keywo ds Machine lea ning ·Sel -adap i e ML ·Sus ainable manu ac u ing ·AI
in manu ac u ing ·O e i ing p e en ion ·Au onomic compu ing
R. Angos o A igues · A. Fe nández Ma ínez (B) · A. G ego es Co o
AIMEN Technology Cen e, O Po iño-Pon e ed a, Spain
e-mail: [email p o ec ed]
J. J. To ez He e a
Ibé ica de Aleaciones Lige as (IDALSA), Remolinos (Za agoza), Spain
© The Au ho (s) 2025
M. Fa manba e al. (eds.), Ho izons o AI: E hical Conside a ions and In e disciplina y
Engagemen s, F on ie s o A i icial In elligence, E hics and Mul idisciplina y
Applica ions, h ps://doi.o g/10.1007/978-981-96-7945-4_27
437
438 R. Angos o A igues e al.
27.1 In oduc ion
The manu ac u ing indus y is a key pilla o he Eu opean economy, accoun ing o
up o one qua e o Eu opean Union’s (EU) business economy ne u no e (“Busi-
nesses in he manu ac u ing sec o —s a is ics explained [online]”). Ne e heless,
Eu opean indus y has been hea ily impac ed in he las yea s due o se e al ac o s,
including ising global compe i ion, he COVID-19 pandemic, and geopoli ical c ises
like he Russian agg ession on Uk aine. Ini ia i es such as he Eu opean Indus ial
S a egy (“Eu opean indus ial s a egy—Eu opean Commission [online]”) and he
G een Deal Indus ial Plan (“EUR-Lex - 52023DC0062 - EN - EUR-Lex [online]”)
aim a enhancing he compe i i eness o he Eu opean indus y by suppo ing he
win g een and digi al ansi ioning o he sec o . As a esul o his end, he up ake
o ad anced echnologies has been accele a ed in manu ac u ing en i onmen s in he
las decade (Calza e al. 2024).
In pa allel wi h hese policy d i e s, he e is a g owing emphasis on he ci cula
economy and closed-loop manu ac u ing, whe e esou ces a e sys ema ically eused
and epu posed o educe bo h was e and cos s. This ocus on sus ainabili y ampli ies
he ele ance o inno a i e da a-d i en ools ha can moni o , p edic , and op imize
a ious indus ial p ocesses. Consequen ly, manu ac u ing companies a e inc eas-
ingly explo ing AI o add ess esou ce sca ci y, ene gy consump ion, and ope a-
ional isks. Despi e he abundan ad an ages, he adop ion o AI can be slowed by
unce ain y ega ding model obus ness and eliabili y, which is especially c i ical in
sensi i e manu ac u ing p ocesses whe e e o s can lead o ope a ional down ime o
subop imal esou ce usage.
A i icial In elligence (AI), conside ed he ou h la ges ad anced manu ac u ing
echnology (Calza e al. 2024), has s and ou as key enable s o sus ainable manu-
ac u ing. The expec ed bene i s o implemen ing AI in manu ac u ing ely on i s
capaci y o enhance esou ce op imiza ion and ope a ional e iciency, while also
educing ene gy consump ion and was e gene a ion (Fe nández Ma ínez e al. 2024;
Wal e smann e al. 2021; Willenbache e al. 2018) which is pa icula ly impo an
in indus ial ecycling p ocesses, cha ac e ized by hese in insic ea u es. In he
me al indus y, a c ucial sec o in he EU economy ha ep esen s a ound 4% o he
EU’s GDP (“Economic and s eel ma ke ou look 2023–2024”, 2023), he EU leads
me al ecycling wi h a ecycling a e o 85% o s eel and 75% o aluminum, he
second mos consumed me al p oduc . In e ms o ene gy sa ings, he la e sec o
can sa e up o 95% o he ene gy equi ed o p ima y aluminum p oduc ion (“New
li e o ecycled aluminium— ecycling magazine [online]”). Howe e , indus ies
ace pa icula challenges due he a iabili y and unce ain y o he composi ion and
quali y o he sc aps ha a e used as p ima y eeds ock, wha a ec s he e iciency
o he o e all ecycling p oduc ion (Capuzzi and Timelli 2018). AI can help mi i-
ga ing hese issues by imp o ing he cha ac e iza ion and so ing accu acy o sc aps
(Capuzzi and Timelli 2018), op imizing p ocessing echniques (Aue e al. 2019),
and suppo ing decision-making ac i i ies, among o he s (“De elopmen o decision
27 A Sel -Adap i e ML Pipeline o Sus ainable Manu ac u ing 439
suppo sys em based on machine lea ning and digi al win o aluminium mel ing
u naces”).
None heless, he up ake o AI in indus ial en i onmen s con ends i s own ba ie s,
including da a a ailabili y and quali y, lack o s anda diza ion, and Human-In-The-
Loop (HITL) adop ion challenges (“AI wa ch: AI up ake in Manu ac u ing—Eu o-
pean Commission [online]”). Conside ing he la e poin , ini ia i es o upskilling
and aining he wo k o ce ha e p o en o be essen ial o inc ease he use s buy-
in o he echnology (Aue e al. 2019). Howe e , he inc easing complexi y o
managing ad anced analy ics so wa e commonly hinde s i s adop ion i n manu-
ac u ing en i onmen s. To help use s o e come hese issues, he in eg a ion o
Au onomic Compu ing (AC) me hods o enhance AI sel -managemen and sel -
adap a ion (Fe nández Ma ínez e al. 2024) has eme ged as p omising solu ions,
os e ing b oade accep ance o AI in he sec o . Ensu ing e icien syne gy be ween
ope a o s and AI o exploi he knowledge coming om human expe ise.
This pape p esen s a no el pipeline o sel -adap i e ML wi h he inco po a ion
o an O e i ing Index o ML o enable au onomous, su oga e-based pa ame e
uning, as illus a ed in Fig. 27.1. The model is designed o assis ope a o s in es i-
ma ing he chemical composi ion o sc ap ecipes in aluminum ecycling p ocesses.
The p oposed pipeline is pa o a b oade Decision Suppo Sys em ha combines an
AI Da a Pipeline and an Au onomic Manage , enabling adap i e model adjus men s
ia he igge ing o sel -abili ies based on human eedback and p ocess me ada a.
This wo k speci ically emphasizes he impo ance o obus o e i ing p e en ion
in eal-wo ld manu ac u ing scena ios, whe e challenges such as ou -o -dis ibu ion
da a and limi ed da a a ailabili y equen ly a ise. By le e aging sel -adap i e abili-
ies, he pipeline enhances model gene aliza ion, ensu ing eliable pe o mance unde
dynamic indus ial condi ions. The me hodology is de ailed in Sec ion II, key esul s
a e discussed in Sec ion III, and he inal conclusions a e p esen ed in Sec ion IV.
Fig. 27.1 Gene al amewo k wi h he Su oga e Adap i e Pipeline (SAP) lying a he co e o he
AI Da a Pipeline (AIDP) p o iding sel -adap i e capabili ies wi h he su oga e-based op imiza ion
o he O e i ing Index (OI). The Au onomic Manage igge s he sel -abili ies in he AIDP
440 R. Angos o A igues e al.
27.2 Me hodology
27.2.1 The Challenge o Indus ial Recycling P ocesses
in he Seconda y Sec o
Seconda y p oduc ion is he p ocess o ecycling o e-p ocessing sc ap ma e ials in o
aluable p oduc s, p omo ing a mo e sus ainable, ene gy-e icien manu ac u ing
compa ed o p ima y p oduc ion (“Me al Recycling Fac shee ”). Howe e , sc ap
ma e ials a e usually ha d o cha ac e ize due o i s he e ogenei y in composi ion
and quali y, magni ied by he lack o in o ma ion abou i s sou ces, wha a ec s he
o e all e iciency o he p ocess (Capuzzi and Timelli 2018). In he con ex o he
aluminum sec o , he ansi ion om c i ical aw ma e ials like bauxi e ( aw o e) o
aluminum sc ap also educes he demand o en i onmen al-ha m ul ac i i ies such
as mining, he eby con ibu ing o he p ese a ion o na u al esou ces and educing
he en i onmen al impac (“The aw-ma e ials challenge: How he me als and mining
sec o will be a he co e o enabling he ene gy ansi ion | McKinsey [online]”).
The seconda y aluminum-making p ocess gene ally s a s wi h he ecep ion o
sc aps and o he ma e ials om di e en sou ces and endo s. I needed, hese
incoming ma e ials migh unde go di e en p ocessing echniques o modi y hei
olume and shape in a sepa a e g inding acili y o op imize hei u u e usage. A e
ecep ion, a sampling o incoming ma e ials can help c oss-checking he composi ion
o sc aps o p ope ly so hem be o e hei s o age in he di e en silos. Acco ding
o he planned o de s, ma e ials a e hen selec ed based on hei a ailabili y, cos ,
and he chemical composi ion ob ained du ing he ecep ion sampling. The se o
ma e ials selec ed o he p oduc ion o aluminum is usually e e ed o as ecipe
o ma e ials. To his day, he selec ion o ma e ials is commonly made by ope a-
o s based on human expe ise. When he ini ial sampling o ma e ials is a ailable,
mass-balance es ima ions can also be made abou he composi ion o he p oposed
mix u e o suppo human decision-making, al hough hese samplings b ing hei
own unce ain ies due o he lack o ep esen a ion o he comple e lo and/o incom-
ple e in o ma ion o he p ope cha ac e iza ion o sc aps due o limi ed sensing
me hods. Da a-d i en app oaches s ep up as al e na i e solu ions o imp o e cu en
es ima ions using his o ical da a o lea n abou pa e ns and he beha io o sc aps
om he p ocess i sel .
Once ma e ials a e selec ed, hey a e ans e ed om he silos o he main indus-
ial acili y. Ma e ials a e measu ed by specialized equipmen and cha ged in he
main u naces ollowing a speci ic o de a ending o hei cha ac e is ics. In he
p ima y mel ing s age, sc aps a e mel ed o ob ain an aluminum mix u e wi h a
desi ed composi ion acco ding o he o de s o clien s. Each o de is de ined by a
p oduc code and a quan i y, and each p oduc code has an associa ed no m ha
cons ains he chemical composi ion allowed acco ding o na ional o in e na ional
no ma i e guidelines, e.g., ISO, UNE. Following p ima y mel ing, e ining ope a-
ions a e equen ly made in a seconda y mel ing s age wi h he inco po a ion o pu e
ma e ials (alloys). Chemical analyses a e conduc ed as needed o gua an ee ha he
27 A Sel -Adap i e ML Pipeline o Sus ainable Manu ac u ing 441
inal mix u e mee s he desi ed no m. The p ocess inalizes wi h he molding and
cooling o he mix u e o ob ain he inal aluminum p oduc s. These p oduc s a e
eady o be sen o he inal clien s.
27.2.2 A Sel -Adap i e ML Pipeline In eg a ing O e i ing
P e en ion
ML models ha e p o en o be success ul enhancing sus ainable manu ac u ing by
op imizing he use o esou ces, ene gy-e iciency, and educing was e (Fe nández
Ma ínez e al. 2024; Wal e smann e al. 2021; Willenbache e al. 2018). Howe e ,
he complexi y o managing ML solu ions poses a signi ican challenge o hei
deploymen in eal indus ial scena ios. Implemen ing, debugging, and main aining
complex ML a chi ec u es call o specialized knowledge, and ope a o s equen ly
do no ha e enough ime o aining o handle ad anced ML, hus aising he ba ie
o success ul AI in eg a ion.
The p inciples o Au onomic Compu ing ha e gained a en ion in ecen yea s as a
means o p o ide sma solu ions wi h sel -adap i e o sel -managing capabili ies o
o e come hese challenges. In any sel -managing sys em, ou dis inc p ope ies can
be dis inguished, namely sel -con igu a ion, sel -healing, sel -op imiza ion, and sel -
p o ec ion (Pa asha and Ha i i 2005). Conside ing ML, sel -con igu a ion is c ucial
o acili a e he adjus men o models, including he ine- uning o hei (hype -)
pa ame e s. None heless, enabling his p ope y is no i ial, as he sys em mus
be able no only o ine- une he model, bu also e alua e he adjus men s made o
alida e he upda ed solu ion be o e deploymen .
O e i ing is a common issue in ML whe e a model lea ns no only he unde -
lying pa e ns in he aining da a, bu also he noise, limi ing i s abili y o gene alize
and mos commonly leading o poo pe o mance in new, unseen da a (Mon esinos
López e al. 2022). Howe e , o e i ing migh no always be appa en when e alu-
a ing model pe o mance o mul iple easons. One o hese easons is an insu icien
es da ase size, which o en a ises om limi ed da a a ailabili y. An inadequa e
es se may ail o ep esen he ull complexi y o eal-wo ld scena ios, leading o
misleading pe o mance indica o s. Addi ionally, a es se oo simila in dis ibu ion
wi h he aining se can lead o o e i ed models ha can misleadingly indica e
good pe o mance, e en hough hey migh lack gene aliza ion capabili ies. Ano he
ac o con ibu ing o unde ec ed o e i ing is elying on a single e alua ion me ic
a he han employing a mul i-me ic app oach. Using mul iple me ics b oadens
he e alua ion scope, o e ing insigh s in o he model’s pe o mance om di e en
pe spec i es and educing he isk o d awing incomple e conclusions. I n o de o
add ess he o e i ing p oblem, c oss- alida ion (e.g., K-Fold and Hold-Ou alida-
ion) and egula iza ion echniques (e.g., L2 and L1 egula iza ion) a e adi ionally
inco po a ed in he aining p ocess. Howe e , hese me hods do no always gua an ee
442 R. Angos o A igues e al.
obus pe o mance i he da ase is highly skewed o i ce ain sc ap composi ions
a e unde ep esen ed.
The O e i ing Index (OI) o ML. Recen wo ks ha e in oduced he concep
o he so-called O e i ing Index, pa icula ly o ad anced models such as Neu al
Ne wo ks (NN) (Abu ass 2023). The OI p oposed in Abu ass (2023) ocused on he
e alua ion o he accu acy loss in bo h he aining and alida ion da ase s, inco po-
a ing he epoch numbe as a weigh ed measu e o emphasize he o e i ing occu -
ing in la e aining epochs. In his wo k, we p opose an O e i ing Index o ML
aining p ocedu es based on a holis ic e alua ion o he adi ional Mean Absolu e
E o (MAE). Unlike p io app oaches ha assess only he inal MAE in he ali-
da ion se , he p oposed OI o ML cap u es he e ol ing pe o mance ac oss he
aining and alida ion da ase s h oughou he en i e aining p ocess by analyzing
he di e ences in hei MAE alues, ha is, he gap be ween he aining and alida-
ion MAE cu es. Addi ionally, we in oduce a no el e m o e alua e he impac o
slope ends, e lec ing he a e o change in e o o e ime. To ensu e a comp ehen-
si e and in e p e able measu e o o e i ing, he index is scaled by he o al numbe o
da a samples used du ing he model lea ning p ocess. This scaling ensu es he p ope
penaliza ion o o e i ing, pa icula ly in la e aining s ages. Hence, he no el y
o ou OI o ML lies in he inco po a ion o wo penaliza ion ac o s. The i s
ac o add esses he gap be ween he aining and alida ion e o cu es, while he
second ac o accoun s o slope ends, o ming a mul i-me ic e alua ion app oach
o o e i ing de ec ion and p e en ion.
The OI o ML is desc ibed in Eq. 27.1.
OI =
k−1
i=1
((w1 · E ain,i − E al,i + w2 · P(i)) · Ni+1) + E al,k · Nk(27.1)
whe e
•E ain,i and E al,i ep esen he nega i e mean absolu e e o s o he aining and
alida ion da ase s a he i h poin o he aining p ocess, espec i ely.
•P(i) is a penaliza ion ac o o he ends in e o slopes desc ibed in Eq. 27.2.
•Ni+1 is he numbe o samples used in aining a he (i + 1) h s ep.
•E al,k ep esen s he inal nega i e mean absolu e e o in he alida ion da ase a
he end o he aining p ocess.
•Nk ep esen s he o al numbe o samples in he da ase used by he lea ning
cu e.
•w1 and w2 a e weigh s assigned o he penaliza ion e ms.
The i s e m o he equa ion, w1 · E ain,i − E al,i , add esses he disc epancy
be ween he aining and alida ion e o s a each s ep, e alua ing in his way he gap
be ween bo h lea ning cu es. The second e m is w2 · P(i) whe e P(i) is desc ibed
in mo e de ail in Eq. 27.2, undesi able ends in he e o slopes a e penalized. Bo h
e ms a e scaled by he numbe o samples a he nex s ep Ni+1, e lec ing he
impac o da a olume in he model’s lea ning p og ess. The hi d e m e e s o he
27 A Sel -Adap i e ML Pipeline o Sus ainable Manu ac u ing 443
commonly used alida ion e o a he end o he aining p ocess, scaled by he o al
numbe o samples o ensu e a simila scale wi h espec o he o he wo e ms. The
weigh s w1 and w2 allow o lexibili y in emphasizing he ele ance o pe o mance
disc epancy e sus slope beha io s e sus inal alida ion e o . The penaliza ion
ac o P(i) ha adjus s he ends in e o slopes is desc ibed in Eq. 27.2.
Pi = Si · log(1 + Ni+1) · Fi(27.2)
whe e
•Si measu es he absolu e di e ence be ween he slopes o he alida ion
and aining e o cu es a he i h poin o he aining p ocess, ha is,
slope ain,i − slope al,i .
•The loga i hmic e m log(1 + Ni+1) scales he penal y based on he size o he
aining da ase .
•Fi is a condi ional penal y ac o .
The di e ence be ween slopes, Si, emphasizes how quickly he e o s a e changing
as new da a is in oduced in he model. On he o he hand, he loga i hmic e m,
log(1 + Ni+1), ein o ces he impac o la ge da ase s in he penaliza ion. The condi-
ional penal y ac o Fi adjus s he penal y based on speci ic condi ions obse ed in
he slope ends:
•Fi = 20 i bo h aining and alida ion slopes a e posi i e and he aining slope
exceeds he alida ion slope, indica ing se e e o e i ing.
•Fi = 10 i he aining slope is posi i e bu he alida ion slope is nega i e,
sugges ing he model is lea ning noise o i ele an pa e ns.
•Fi = 2 i only he alida ion slope is posi i e, indica ing ising e o s in alida ion,
a milde o m o o e i ing.
Su oga e Models o Op imizing Pa ame e Explo a ion. Su oga e models,
also known as me amodels, a e simpli ied models ha app oxima e mo e complex
unde lying unc ions o ML models. They a e pa icula ly use ul in scena ios whe e
e alua ion o he o iginal model is compu a ionally expensi e o ime-consuming.
By i ing a su oga e model o he da a de i ed om he o iginal model’s pe o -
mance ac oss a ious con igu a ions, i is possible o e icien ly explo e pa ame e
spaces and p edic he pe o mance o new con igu a ions wi h signi ican ly educed
compu a ional o e head.
Inco po a ing su oga e models in o he ine- uning p ocess enables a mo e e i-
cien and sys ema ic app oach o iden i ying op imal model adjus men s. By le e -
aging he OI as he objec i e unc ion, he ine- uning app oach accele a es he
iden i ica ion o op imal model adjus men s while minimizing compu a ional and
esou ce cos s (Ba wey e al. 2023). The p oposed me hodology suppo s a sel -
adap i e pipeline capable o au onomously adjus ing model pa ame e s in esponse
o changes in da a o pe o mance ends obse ed in eal-wo ld scena ios. This adap -
abili y ensu es con inuous model op imiza ion wi h minimal o no use in e en ion,
enhancing bo h e iciency and obus ness.
444 R. Angos o A igues e al.
Fig. 27.2 Model pipeline componen o he sel -adap i e pipeline (SAP) o machine lea ning
The Sel -Adap i e Pipeline (SAP) o ML. The SAP p oposed in his wo k is
composed o h ee main sub-pipelines componen s. A he co e o SAP, a Model
Pipeline was de ined o p e-p ocess aw da a ia a Column T ans o me componen
ollowed by he de ini ion o a Mul i-Ou pu Reg esso (MOR) model as shown in
Fig. 27.2. Column T ans o me accoun s o he ype o da a by encoding ca ego ical
a iables wi h me hods like one-ho encode , while scaling nume ical inpu s o ensu e
uni o mi y be ween pa ame e s. The MOR is w apped a ound a Random Fo es
Reg esso model capable o p edic ing mul iple a ge a iables om a se o inpu s.
This se up i s essen ial o handling he complex ela ionships o en ound in ma e ial
composi ion da a.
The second block o he SAP co esponds o he Su oga e Model Op imiza ion
Pipeline, shown in he bo om sec ion o Fig. 27.3, which is essen ial o educe
compu a ional o e head du ing he hype pa ame e uning phase. In his s age, he
objec i e unc ion p e iously de ined and he pa ame e space o ine- uning a e ed
o a loop in which a Bayesian op imiza ion is conduc ed using he OI o ML as
key me ic. Following his app oach, he sea ch o hype pa ame e s is guided owa d
con igu a ions ha balance model accu acy and gene aliza ion. Each i e a ion o he
op imiza ion in ol es e alua ing po en ial hype pa ame e s by unning he su oga e
model, which p edic s he OI based on his o ical da a om p e ious e alua ions.
Based on his p ocedu e, i is ensu ed ha only he mos p omising con igu a ions
a e ully e alua ed using he mo e compu a ionally in ensi e p ima y model, he eby
s eamlining he en i e model aining p ocess. In he las s age o he SAP, he bes
se o hype pa ame e s iden i ied is selec ed o ain he p ima y model on he ull
da ase o e i y he pe o mance me ics sugges ed by he su oga e model, as shown
in he main pipeline a he uppe sec ion o Fig. 27.3.
27 A Sel -Adap i e ML Pipeline o Sus ainable Manu ac u ing 445
Fig. 27.3 Wo k low o he su oga e model op imiza ion pipeline in he sel -adap i e pipeline
(SAP) o machine lea ning
27.2.3 A Me ada a-D i en Au onomic Manage o T igge
Sel -abili ies on an AI Da a Pipeline
The p oposed me hodology encapsula es a sel -adap i e mechanism whe eby he
model can dynamically adjus i s pa ame e s in esponse o shi s in da a cha ac-
e is ics o pe o mance objec i es. This SAP is pa o a la ge AI Da a Pipeline
(AIDP) ha comp ises i e componen s o enable he de elopmen and deploymen
o ML models in eal indus ial applica ions (A igues e al. 2024). These componen s
include Da a Inges ion, Da a T ans o ma ion, Da a Explo a ion, Model T aining,
and Real-Wo ld Usage. Th oughou hese s ages o he AIDP, a dedica ed Me a-
da a Logging module eco ds impo an s a is ics. Me ada a, as i s name sugges s,
co esponds o da a om da a as a highe -le el o abs ac ion o he componen s pe
se. Some o his me ada a include human eedback abou he pe o mance o he
ML models and impo an s a is ical in o ma ion o new da a (e.g., da a dis ibu ion
shi s, use o e ides, o anomalies in daily sc ap deli e ies). An ex e nal Au onomic
Manage componen moni o s he me ada a om he AIDP, analyzes i , plans abou
he co ec i e measu emen s needed, and execu es hem back in he AIDP h ough
he ac i a ion o ala ms and igge ing o he sel -abili ies ollowing he MAPE-K
a chi ec u e. Because his p ocess is e en -d i en, i allows he en i e pipeline o
emain do man when e e y hing is s able, hus a oiding unnecessa y compu a ions
and sys em o e head.
452 R. Angos o A igues e al.
o con igu a ions wi h highe OIs h ough his su oga e model app oach add signi i-
can alue o he uning p ocess, op imizing model pe o mance mo e e icien ly and
ensu ing i s obus ness unde a ying ope a ional condi ions.
As pa o a bigge amewo k ha igge s he ac i a ion o au onomous abili ies
based on eal- ime human eedback and moni o ed da a, his me hodology se s a
new s anda d in machine lea ning op imiza ion. I s abili y o e icien ly ecommend
model adjus men s ensu es op imal pe o mance while educing eliance on ex ensi e
manual in e en ion. The app oach no only ad ances ML op imiza ion bu also o e s
a scalable solu ion o indus ies seeking o s eamline hei ML model de elopmen
li ecycle.
Fu he e inemen s and enhancemen s o he O e i ing Index me hodology will
be explo ed in u u e wo k o op imize i s applicabili y and e iciency in di e se
modeling scena ios. Re inemen s o he OI me ic—such as adap i e penaliza ion
ac o s ha become s ic e o e ime—could u he shield p oduc ion lines om
pe o mance deg ada ion in scena ios o apid da a d i . Besides, in eg a ing hese
solu ions in o b oade en e p ise sys ems emains a ui ul di ec ion o u u e
esea ch. Addi ional in es iga ion in o mul i- ask lea ning, domain adap a ion, o
ad anced sampling o unde ep esen ed sc ap ypes can u he en ich he pipeline’s
applicabili y in he u u e.
Acknowledgemen s This wo k has been suppo ed by he p ojec ‘sel -X A i icial In elligence
o Eu opean P ocess Indus y digi al ans o ma ion’ (s-X-AIPI), which has ecei ed unding om
he Eu opean Union’s Ho izon Eu ope esea ch and inno a ion p og am unde G an Ag eemen
No. 101058715.
Re e ences
7.2. Real wo ld da ase s—sciki -lea n 1.5.1 documen a ion [online]. A ailable: h ps://sciki -lea n.
o g/s able/da ase s/ eal_wo ld.h ml#cali o nia-housing-da ase . Accessed 6 Aug 2024
Abu ass S (2023, Aug) Quan i ying o e i ing: in oducing he O e i ing Index [online].
A ailable: h ps://a xi .o g/abs/2308.08682 1.Accessed 5 Aug 2024
AI wa ch: AI up ake in Manu ac u ing—Eu opean Commission [online]. A ailable: h ps://ai-
wa ch.ec.eu opa.eu/publica ions/ai-wa ch-ai-up ake-manu ac u ing_en. Accessed: 7 Aug 2024
A igues R, Co o A, He e a J, Tomás F, Ve a di S, Ma zano M, Ma inez A (2024) An AI-d i en
use -cen ic amewo k ein o ced by au onomic compu ing: A case s udy in he Aluminium
sec o . In: Ta eq Ah am, Luca Casa o o and Pie o Cos a (eds) Human in e ac ion and eme ging
echnologies (IHIET 2024). AHFE (2024) In e na ional Con e ence. AHFE Open Access, ol
1. AHFE In e na ional, USA. h ps://doi.o g/10.54941/ah e1005478
Aue M, Osswald K, Volz R, Woidasky J (2019, Ma ) A i icial in elligence-based p ocess o me al
sc ap so ing [online]. A ailable: h ps://a xi .o g/abs/1903.09415 1. Accessed 5 Aug 2024
Ba wey S, Kim H, Maulik R (2023, No ). In e p e able ine- uning and e o indica ion o
g aph neu al ne wo k su oga e models [online]. A ailable: h ps://a xi .o g/abs/2311.07548 3.
Accessed 5 Aug 2024
Businesses in he manu ac u ing sec o —s a is ics explained [online]. A ailable: h ps://ec.eu
opa.eu/eu os a /s a is ics-explained/index.php? i le=Businesses_in_ he_manu ac u ing_sec o .
Accessed 15 Jul 2024
27 A Sel -Adap i e ML Pipeline o Sus ainable Manu ac u ing 453
Cali o nia housing p ice p edic ions [online]. A ailable: h ps://www.kaggle.com/code/eswa c
hand /cali o nia-housing-p ice-p edic ions. Accessed 6 Aug 2024
Calza E, Sogue o EJ, Fabiani J, De PG (2024, Jun) Ad anced manu ac u ing s udy. P elimina y
indings on EU’s ad anced manu ac u ing indus y in he global landscape. Jun, h ps://doi.o g/
10.2760/798090
Capuzzi S, Timelli G (2018) Ap ) P epa a ion and mel ing o sc ap in aluminum ecycling: a e iew.
Me als 8(4):249. h ps://doi.o g/10.3390/MET8040249
De elopmen o decision suppo sys em based on machine lea ning and digi al win o aluminium
mel ing u naces. h ps://doi.o g/10.5281/ZENODO.5897240
Economic and s eel ma ke ou look 2023–2024 Fi s qua e epo Da a up o, and including, hi d
qua e 2022 In oduc ion, 2023
EUR-Lex - 52023DC0062 - EN - EUR-Lex [online]. A ailable: h ps://eu -lex.eu opa.eu/legal-con
en /EN/TXT/?u i=CELEX%3A52023DC0062. Accessed 5 Aug 2024
Eu opean indus ial s a egy—Eu opean Commission [online]. A ailable: h ps://commission.eu
opa.eu/s a egy-and-policy/p io i ies-2019-2024/eu ope- i -digi al-age/eu opean-indus ial-s
a egy_en. Accessed 7 Aug 2024
Fe nández Ma ínez A e al (2024) A machine lea ning amewo k o imp o ing esou ces, p ocess,
and ene gy e iciency owa ds a sus ainable s eel indus y. LNNS 1028:3–10. h ps://doi.o g/10.
1007/978-3-031-61905-2_1
Mon esinos López OA, Mon esinos López A, C ossa J (2022) O e i ing, model uning, and e alua-
ion o p edic ion pe o mance. Mul i a ia e S a is ical Machine Lea ning Me hods o Genomic
P edic ion, pp 109–139. h ps://doi.o g/10.1007/978-3-030-89010-0_4
New li e o ecycled aluminium— ecycling magazine [online]. A ailable: h ps://www. ecycling-
magazine.com/2021/09/02/new-li e- o - ecycled-aluminium/. Accessed 5 Aug 2024
“Me al Recycling Fac shee ”.
Pa asha M, Ha i i S (2005) Au onomic compu ing: an o e iew. LNCS 3566:257–269. h ps://doi.
o g/10.1007/11527800_20
The aw-ma e ials challenge: How he me als and mining sec o will be a he co e o enabling he
ene gy ansi ion | McKinsey [online]. A ailable: h ps://www.mckinsey.com/indus ies/me
als-and-mining/ou -insigh s/ he- aw-ma e ials-challenge-how- he-me als-and-mining-sec o -
will-be-a - he-co e-o -enabling- he-ene gy- ansi ion. Accessed 5 Aug 2024
Wal e smann L, Kiemel S, S uhlsa z J, Saue A, Miehe R (2021) Jun) A i icial in elligence applica-
ions o inc easing esou ce e iciency in manu ac u ing companies—A comp ehensi e e iew.
Sus ainabili y 13(12):6689. h ps://doi.o g/10.3390/SU13126689
Willenbache M, Kunisch C, Wohlgemu h V (2018) Applica ion o me hods o a i icial in elligence
o sus ainable p oduc ion o manu ac u ing companies, pp 225–236. h ps://doi.o g/10.1007/
978-3-319-65687-8_20
XGBoos o he Cali o nia housing da ase | XGBoos ing [online]. A ailable: h ps://xgboos ing.
com/xgboos - o - he-cali o nia-housing-da ase /. Accessed 7 Aug 2024
454 R. Angos o A igues e al.
Open Access This chap e is licensed unde he e ms o he C ea i e Commons A ibu ion-
NonComme cial-NoDe i a i es 4.0 In e na ional License (h p://c ea i ecommons.o g/licenses/by-
nc-nd/4.0/), which pe mi s any noncomme cial use, sha ing, dis ibu ion and ep oduc ion in any
medium o o ma , as long as you gi e app op ia e c edi o he o iginal au ho (s) and he sou ce,
p o ide a link o he C ea i e Commons license and indica e i you modi ied he licensed ma e ial.
You do no ha e pe mission unde his license o sha e adap ed ma e ial de i ed om his chap e
o pa s o i .
The images o o he hi d pa y ma e ial in his chap e a e included in he chap e ’s C ea i e
Commons license, unless indica ed o he wise in a c edi line o he ma e ial. I ma e ial is no
included in he chap e ’s C ea i e Commons license and you in ended use is no pe mi ed by
s a u o y egula ion o exceeds he pe mi ed use, you will need o ob ain pe mission di ec ly om
he copy igh holde .
